BangInput 开源资产镜像

# BangInput Open Assets Mirror This dataset repository is the mainland China download mirror for redistributable BangInput assets. Only assets whose upstream license and redistribution terms have been reviewed are published here. The first repository is restricted to Apache-2.0-compatible assets. Every immutable object is stored under a versioned or SHA-256-addressed path and is checked against the digest registered by the BangInput client. ## Object layout ```text models/asr/{modelId}/{sha16-or-upstream}/{archive} models/llm/{modelId}/{sha16}/{file} models/association/{modelId}/{sha16}/{file} models/lm/{version}/{sha16}/{file} dicts/cell/{sha16}/{file} fonts/v1/{sha16}/{file} catalogs/cell/catalog.json ``` This repository contains no user input, telemetry, personal data, APK signing material, or service credentials. ## Published assets - `predictive-text-small`: ONNX model and vocabulary for offline next-word prediction. - `Qwen3-0.6B-IQ4_XS`: pinned GGUF used by the offline correction pipeline. - `X-ASR`: pinned streaming and offline Chinese/English ASR archives. - `Qwen3-ASR 0.6B int8`: pinned multilingual offline ASR archive. The machine-readable source, size, SHA-256 and object path for every published file are recorded in `assets.json`. Assets with unresolved, non-commercial or incompatible terms are intentionally excluded. MIT, CC BY 4.0, and custom-license weights are kept in separate dataset repositories; their manifests and license notices live in the corresponding subdirectories of this folder. ## License-separated repositories - Apache-2.0: `JuneLegency/banginput-assets-apache2` - MIT: `JuneLegency/banginput-assets-mit` - CC BY 4.0: `JuneLegency/banginput-assets-cc-by-4` - FunASR Model Open Source License 1.1: `JuneLegency/banginput-assets-custom-license` TeleSpeech is not mirrored: its model license requires prior written authorization for commercial use. Project: https://github.com/JuneLeGency/BangInput